Autofluorescence Bronchoscopy Video Analysis for Lesion Frame Detection^.

Area of Science:

• Pulmonology
• Medical Imaging
• Artificial Intelligence

Background:

• Bronchial lesions are critical indicators of early lung cancer and squamous cell carcinoma.
• Autofluorescence bronchoscopy (AFB) is sensitive for detecting suspicious lesions but requires tedious manual video review.
• Automated analysis of AFB videos is underexplored for efficient lesion detection.

Purpose of the Study:

• To develop a robust automated approach for analyzing AFB videos.
• To distinguish informative from uninformative video frames.
• To identify and delineate potential bronchial lesions within informative frames.

Main Methods:

• Utilized a combination of computer-based image analysis, machine learning, and deep learning.
• Developed an algorithm to classify AFB video frames as informative or uninformative.
• Implemented lesion detection and region delineation for informative frames.

Main Results:

• Achieved 99.5% accuracy in labeling informative frames and 90.2% for uninformative frames.
• Outperformed ResNet in frame classification accuracy (99.2%/47.6%).
• Correctly identified ≥97% of lesion frames with ≤3% false positive/negative rates.

Conclusions:

• The proposed automated AFB analysis method significantly enhances efficiency in bronchial lesion detection.
• This approach aids in advancing early lung cancer detection by making AFB video analysis more tractable.
• The findings support the clinical relevance of AI-driven tools for improving diagnostic accuracy in pulmonology.